Skip to content

A Neural Machine Translation toolkit for research purpose

License

Notifications You must be signed in to change notification settings

quanpn90/NMTGMinor

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Introduction

Requirements and Installation

Currently NMTG requires PyTorch version >= 1.8.0. Best is 1.10.0 Please follow the instructions here: https://github.com/pytorch/pytorch#installation.

After PyTorch is installed, you can install the requirements with:

pip install -r requirements.txt

C++/CUDA module installation

NMTG supports a couple of modules written using custom Pytorch/C++/CUDA modules to utilize GPU better and reduce overheads, including:

  • Self-attention and encoder-decoder attention with CUBLASLT
  • Multi-layer Perceptrons with CUBLASLT and fused dropout-relu/gelu/silu where inplace is implemented whenever possible
  • Highly optimized layer norm and multi-head attention (only available with sm80 (NVIDIA A100)) from Apex
  • Fused Logsoftmax/Cross-entropy loss to save memory for large output layer, from Apex
  • Fused inplaced Dropout Add for residual Transformers

Installation requires CUDA and nvcc with the same version with PyTorch. Its possible to install CUDA from conda via:

conda install -c nvidia/label/cuda-11.3.1 cuda-toolkit

or if using a custom version with CUDA 11.5

conda install -c nvidia/label/cuda-11.5.2 cuda-toolkit

(depending on the CUDA version that comes with your PyTorch)

And then navigate to the extension modules and install nmtgminor-cuda via

cd onmt/modules/extension
python setup.py install

Without this step, all modules backoff to PyTorch versions.

IWSLT 2022 Speech Translation models

Interspeech 2022 Multilingual ASR models

About

A Neural Machine Translation toolkit for research purpose

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published